KR102011453B1 - Predictability of Conditionality Using Recurrent Neural Networks - Google Patents

Abstract

Methods, systems, and apparatus are disclosed that include a computer program encoded on a computer storage medium for predicting the likelihood of conditions met using a recurrent neural network. One of the systems comprises one or more concurrent neural network layers configured to process a temporal sequence comprising respective inputs in each of a plurality of time steps; And one or more logistic regression nodes, wherein each of the logistic regression nodes corresponds to a respective condition of a set of predetermined conditions, each of the logistic regression nodes for each of the plurality of time steps; Process the network internal state for the time step according to the current values of the set of parameters of the logistic regression node to receive a network internal state for and to generate a future condition score for the corresponding condition for the time step. do.

Description

Predictability of Conditionality Using Recurrent Neural Networks

The present invention is directed to processing temporal sequences using current neural networks.

A neural network is a machine learning model that uses one or more layers of nonlinear units to predict the output on a received input. Some neural networks include one or more hidden layers in addition to the output layer. The output of each of the hidden layers is used as input to the next layer of the network (ie, the next hidden layer or output layer). Each layer of the network generates an output from the received input, according to the current values of each parameter set.

Some neural networks are recurrent neural networks. Recurrent neural networks are neural networks that receive an input sequence and generate an output sequence from the input sequence. In particular, the recurrent neural network may use all or part of the network internal state at the previous time step in calculating the output at the current time step.

In general, one innovative aspect of the invention described herein utilizes a recurrent neural network system that includes one or more recurrent neural network layers and one or more logistic regression nodes, each of which in each of a plurality of time steps. And a method for processing a temporal sequence comprising an input, the method comprising: for each of the plurality of time steps, receiving the input at the time step; Processing the input at the time step via one or more concurrent neural network layers to generate a network internal state for the time step; And processing the network internal state using each of the one or more logistic regression nodes, wherein each of the logistic regression nodes corresponds to a respective condition of a set of predetermined conditions, and the one or more logistic regression nodes. The processing of the network internal state using each of the following steps includes: generating the logistic regression node according to current values of the set of parameters of the logistic regression node to generate a future condition score for the corresponding condition for the time step. Processing the network internal state for the time step, wherein a future condition score for the corresponding condition indicates the likelihood that the corresponding condition is met within a particular time period of the input in the time step. .

Other embodiments of this aspect include computer programs recorded on corresponding computer systems, apparatuses, and one or more computer storage devices, each of which is configured to perform the operations of the methods.

A system of one or more computers may be configured to perform specific operations or operations by having software, firmware, hardware, or a combination thereof, installed on the system that, in operation, causes the system to perform these operations. . One or more computer programs may be configured to perform particular operations or operations by including instructions that cause the apparatus to perform operations when executed by a data processing apparatus.

The foregoing and other embodiments may each optionally include one or more of the following features, alone or in combination.

The recurrent neural network system may further comprise an output layer. For each of the plurality of time steps, the network internal state for the time step uses the output layer according to the current values of the set of parameters of the output layer to generate each of the next input scores for each of the plurality of possible next inputs. Can be processed. Each of the next input scores for each of the plurality of possible next inputs may indicate the likelihood that the possible next input will be the input at the next time step of the time sequence. Each of the one or more concurrent neural network layers is configured to: for each of the plurality of time steps: receive a layer input for a time step; And, in order to generate a layer internal state for the recurrent neural network layer for the time step, according to the current values of the set of parameters for the recurrent neural network layer, in-layer for the recurrent neural network layer for the preceding time step. And process hierarchical input for the state and time step. A recurrent neural network system can include only one recurrent neural network layer. The layer internal state for the recurrent neural network layer for the time step may be the network internal state for the time step. The recurrent neural network system may include a plurality of recurrent neural network layers arranged in a sequence. The network internal state for a time step may be the layer internal state of the highest relative neural network layer of the sequence for the time step. The recurrent neural network system may include a plurality of recurrent neural network layers. The network internal state for a time step may be a combination of hierarchical internal states of the concurrent neural network layers for the time step. Each input in the temporal sequence may be selected from a set of possible inputs for the recurrent neural network system. The set of predetermined conditions may include conditions that are met if events that are not included in the set of possible inputs to the recurrent neural network system occur within a certain time period of input at the time step. Each of the inputs in the plurality of time steps may be health-related data associated with the patient. The conditions within the predetermined set of conditions may be conditions met when a health-related event occurs in relation to the patient.

Particular embodiments of the invention described herein may be implemented to realize one or more of the following advantages. Recurrent neural networks can be efficiently used to analyze a sequence of events, such as health events, eg, a sequence of health events derived from an electronic medical record for the current patient. A recursive neural network can be used effectively to predict the likelihood that events will occur within a particular time period of the most recent event in a temporal sequence, even if the events are not included in the set of possible inputs to the recursive neural network. Can be. Recurrent neural network internal states can be effectively used to identify different temporal sequences corresponding to different patients, which can include health events that are predictions for future health events that may be associated with the current patient. . Accordingly, embodiments of the present invention provide an improved system and method for processing temporal sequences.

Information characterizing the output of the recurrent neural network or outputs derived from the output generated by the recurrent neural network can be provided to end users, such as doctors or other healthcare professionals, to provide patients with quality medical care. Improve the ability of healthcare professionals to offer. For example, a healthcare professional may use useful information about future health events that may be associated with the current patient, such as health events that may be the next health event associated with the patient or a specific time period of the most recent event in the sequence. Events that occur within can be provided with useful information about the likelihood that certain events will be met. In addition, the healthcare medical professional may be provided with information identifying the potential effect of the proposed treatment on the likelihood of events occurring, e.g., the proposed treatment may result in health-related conditions that are undesirable for the patient in the future. It may be provided whether to reduce or increase the likelihood of meeting. In addition, the healthcare professional may be provided with the healthcare records of patients whose healthcare records were similar to the current patient at one point in the past history or may be provided with a summary of the healthcare results of these patients. In addition, in some cases, a warning may be issued to a healthcare professional that is triggered if the suggested action to be taken by the healthcare professional significantly increases the risk for future predicted outcomes of that patient. In addition, a healthcare analysis system comprising a recurrent neural network can codify standard medical behavior, discover patterns in treatment and outcomes, analyze existing medical technologies or healthcare systems, or It can be used to make recommendations or to promote scientific findings.

The details of one or more embodiments of the invention herein will be set forth in the accompanying drawings and the description below. Other features, aspects, and advantages of the subject matter of the present invention will become apparent from the description, the drawings, and the claims.

1 illustrates an example of a healthcare analysis system.
2 is a flow diagram of an example process for generating health event data for a temporal sequence.
3 is a flow diagram of an example process for generating health analysis data for a time sequence from subsequent input scores.
4 is a flow diagram of an example process for generating health event data for a time sequence from a network internal state.
5 is a flowchart of an example process for generating health event data for a time sequence from future condition scores.
6 is a flowchart of an example process for determining the effect of adding an event to a time sequence for future condition scores.
7 illustrates an example of a recurrent neural network configured to generate a future condition score.
8 is a flow diagram of an example process for generating a future condition score for a given time step.
9 is a flow diagram of an example process for training a reactive neural network to generate a future condition score.
Like reference numbers and designations in the various drawings indicate like elements.

The present disclosure generally relates to health analysis data from a temporal sequence comprising data identifying a plurality of events, such as health events using a recurrent neural network. Describe a system that can generate analytical data, such as health analysis data.

1 shows an example healthcare analysis system 100. Healthcare analysis system 100 is an example of a system implemented as computer programs on one or more computers at one or more locations in which the systems, components, and techniques described below may be implemented.

The healthcare analysis system 100 receives the temporal sequences and processes the received temporal sequences using the recurrent neural network 110 to thereby analyze the health from the received temporal sequences. Generate health analysis data. For example, healthcare analysis system 100 may receive temporal sequence 102, and generate health analysis data 122 from temporal sequence 102.

Temporal sequences are sequences that include health-related data, eg, data identifying a health event, in each of a plurality of time steps. Each temporal sequence includes health-related data associated with a given patient, where the health events identified by the health-related data in the temporal sequence are sorted in chronological order, such that the most recently occurring health event is sequenced. This is the health event at the last type step in the.

In some implementations, temporal sequence generation system 104 generates temporal sequence 102 from an electronic medical record for the corresponding patient. An electronic medical record is an electronic collection of health information for the corresponding patient. For example, a temporal sequence generation system can obtain an electronic medical record for a patient from an electronic medical record repository 106, and identify health events in the electronic medical record and By sorting in chronological order, a temporal sequence 102 can be generated from the electronic medical record. In particular, temporal sequence 102 may include a sequence of tokens in each of the plurality of time steps, where each token represents a health event identified in an electronic medical record. In some implementations, the temporal sequence generation system can append data identifying the time at which the health event occurred to data identifying the health event in temporal sequence 102.

In general, health events identified within the temporal sequences received by the healthcare analysis system 100 may include symptoms, tests, test results, diagnostics, and medications. one or more of medications, outcomes, and the like, each of which is represented by a token from a pre-determined vocabulary of tokens. Optionally, each token is combined with data identifying the time at which the health event occurred within the temporal sequence. In addition, in some cases, the temporal sequence may identify health events other than health events identified by tokens from the vocabulary. For example, in some implementations, health events in temporal sequences can also be health-related images (eg, X-Ray or other diagnostic images). images)), health-related electronic documents (eg, free-form notes generated by a physician during an appointment), or both. .

In addition, optionally, the health-related data may include other health-related data that may be classified as affecting the patient's health. For example, other data may include data characterizing the patient's activity, or collected by activity tracking applications running on the patient's devices (eg, activity tracking devices) or mobile devices. Other health-related data. For example, activity data may include data identifying the distance a patient traveled on a particular date, a workout or other fitness activity in which the patient participated, the foods the patient ate, and the like. Can be. Other health-related data may also identify other data that may be considered to affect the patient's health, such as prescription fulfillment data for the patient or purchases made by the patient. May contain data.

The healthcare analysis system 100 processes the temporal sequence 102 using the recurrent neural network 110 to generate a network output for the temporal sequence 102. The healthcare analysis system 100 also includes a healthcare analysis engine 120, which receives the network output for the temporal sequence 102, and the network Generate output analysis data 122 for temporal sequence 102 from the output.

In general, the network output for the temporal sequence 102 is a set of next input scores 112 of the recurrent neural network 110, future condition scores ( 114, or one or more of network internal state 116.

Recurrent neural network 110 includes one or more recurrent neural network layers that generate a network internal state for each time step of a given input temporal sequence. In some implementations, the recurrent neural network 110 also includes an output layer, or a set of logistic regression nodes, or both, which receive network internal state. And process the network internal state to generate a network output for the time step. Additionally, in some implementations, the recurrent neural network can also include one or more other types of neural network layers, eg, forward layers, such as a fully-connected layer. , Convolutional layers, pooling layers, regulation layers, and the like.

In particular, each of the recurrent neural network layers is configured to receive a layer input for a time step and compute a layer internal state for the layer for the time step. The recurrent neural network layer is a layer internal state for the current time step from the layer input for the current time step and the layer internal state of the previous time step according to the current values of the set of layer parameters. Calculate In some implementations, one or more of the recurrent neural network layers are configured to also use other internal states when calculating the layer internal state for a time step, eg, in a layer from other previous time steps. Internal states for the current time step or previous time steps for the other concurrent layers. If the current time step is the first time step in the sequence, the hierarchical internal state for the previous time step may be, for example, as specified by the system administrator, or the healthcare analysis system 100. As generated by, the initial layer internal state.

If there is only one concurrent neural network layer in the recurrent neural network 110, then the network internal state for a given time step is the layer internal state for the concurrent neural network layer for the time step.

If there are multiple concurrent neural network layers in the recurrent neural network 110, the layers are arranged in a sequence, from the lowest layer in the sequence to the highest layer in the sequence, and within the network for the time step. The health event is collectively processed in the time step to calculate the state. If there are other types of neural network layers in the recurrent neural network 110, the other neural network layers may be intersperated at various locations in the sequence, eg, before the first recurrent layer, Or may be placed between two separate layers, or after all the concurrent layers, or in any combination thereof. For a given time step, the recurrent neural network 110 may provide the layer internal state from each of the recurrent neural network layers as a layer input to the recurrent neural network layer above the layer in the sequence. In some implementations, one or more of the recurrent neural network layers are also configured to receive inputs from one or more other layers in such a sequence that are different from the layer below the recurrent layer.

In some implementations, one or more of the layers in the sequence are part of a layer input to the layer in a subset of time steps, eg, in the first time step or in each time step. , Global input, or per-record input, or both. Global inputs are inputs that do not depend on the current temporal sequence being processed by the recurrent neural network 110. An example of a global input is data characterizing the current time of year, for example the current date. Record-by-record inputs are inputs that may be different for different temporal sequences. Examples of record-by-record inputs may include the patient's genetic sequence associated with the current temporal sequence or other information characterizing the patient, such as demographic information about the patient.

In some implementations, if there are a plurality of concurrent neural network layers, the network internal state for the time step is the highest layer inner state in the sequence for the time step. In some other implementations, the healthcare analysis system 100 combines the hierarchical internal states for the time step to generate a network internal state for the time step. For example, healthcare analysis system 100 may calculate the sum, product, or average of intra-layer states, or concatenate intra-layer states, to generate a network internal state. You can.

In some implementations, the concurrent neural network layers are Long Short-Term Memory (LSTM) layers. Each LSTM layer includes one or more LSTM memory blocks. Each LSTM memory block may comprise one or more cells, each cell comprising an input gate, a forget gate, and an output gate, which are cells It is possible to store previous states for the cell, for example to store previous states for the cell for use in generating current activation or to be provided to other components of the LSTM neural network. To be able.

In implementations where the recursive neural network 110 includes an output layer, for each of the time steps, the output layer receives the network internal state for the time step and sets the next set of input scores for the time step. It is configured to generate. The next set of input scores for the time step includes a respective score for each health event represented by the token within the vocabulary of tokens. When the recurrent neural network 110 is trained, the next input score for a given health event indicates the likelihood that the health event will be the next health event in a temporal sequence. Thus, when the recurrent neural network 110 includes an output layer, for each time step of a given input temporal sequence, the future neural network 110, the future time events, i.e., the next time in the temporal sequence. A network trained to predict health events at a step. Recurrent neural network 110 may be trained on training sequences using conventional machine learning training techniques, eg, backpropagation via a time training technique.

In such implementations, the next input scores 112 for the temporal sequence 102 are the next input scores generated by the output layer for the last time step in the temporal sequence 102.

In implementations in which the recursive neural network 110 includes a set of logistic regression nodes, the set of logistic regression nodes receives, at each time step, a network internal state for the time step, and for the time step. And generate a set of future condition scores. The set of future condition scores includes a respective score for each condition in a pre-determined set of conditions. The score for a given condition indicates the likelihood that the condition is met within the specified duration of the health event at the current time step.

Conditions include conditions that are met by the occurrence of an event, for example, conditions that are met by the occurrence of a health event represented by a token in the vocabulary. In some cases, in addition to or instead of including the conditions met by the occurrence of an event represented by a token in the vocabulary, the conditions in the predetermined set of conditions may also be events not represented by the tokens in the vocabulary. That is, events that are not possible health events included in the temporal sequences processed by the recurrent neural network 110 may include conditions that are met when they occur within a specific time period of the health event at the current time step. Thus, events that can satisfy conditions within a predetermined set of conditions may overlap with the events represented by tokens, but the set of conditions is also caused by the occurrence of other conditions not present in the set. It may include conditions that are met.

A recurrent neural network comprising a set of logistic regression nodes is described in more detail with reference to FIGS. 7 and 8. Training the recurrent neural network to predict the likelihood that the conditions will be met is described in more detail below with reference to FIG. 9.

In such implementations, the condition scores 114 for the temporal sequence 102 are future condition scores generated by logistic regression nodes for the last time step in the temporal sequence 102.

In embodiments where the network internal state 116 is included in the network output for the temporal sequence 102, the network internal state 116 for the temporal sequence 102 is a recursive neural network for the last time step in the sequence. A combination of network internal states generated by the recurrent neural network 110 for a plurality of time steps in the sequence, or generated by the 110, for example, of network internal states. It is a weighted sum, product, or concatenation.

The healthcare analysis engine 120 receives the network output for the temporal sequence 122, generates the healthcare analysis data 122 for the temporal sequence 102, and sends the user, eg, a temporal sequence. Health analysis data 122 is provided for presentation to a doctor treating a patient corresponding to 102. In general, health analysis data 122 is data characterizing future events that may be associated with temporal sequence 102, ie, health events or other events that may occur after the current last health event in temporal sequence 102. To characterize the data.

In implementations in which the neural network output for temporal sequence 102 includes the following input scores 112, healthcare analysis engine 120 identifies health events that may occur next within temporal sequence 102. Health analysis data 122 is generated. Generating health analysis data for the temporal sequence from the following input scores is described in more detail below with reference to FIG. 3.

In implementations in which the neural network output for temporal sequence 102 includes network internal state 116, healthcare analysis engine 120 is more likely to predict future events within temporal sequence 102. Health analysis data 122 that identifies health events from the data sets. In particular, the healthcare analysis engine 120 identifies internal states similar to the network internal state 116 from internal states stored in the internal state repository 130, and is likely to predict future events within the temporal sequence 102. Use these similar internal states to determine health events from other high temporal sequences. The internal state repository 130 stores network internal states generated at various time steps in various temporal sequences, and stores each network internal state as a time step and temporal target when the network internal state occurs. Associate with the data identifying the sequence. Generating health analysis data for the temporal sequence from the network internal state is described in more detail below with reference to FIG. 4.

In embodiments where the neural network output for temporal sequence 102 includes future condition scores 114, healthcare analysis engine 120 generates health analysis data 122 that characterizes scores for conditions. Let's do it. Generating health analysis data for a temporal sequence from future health condition scores is described in more detail below with reference to FIG. 5.

2 is a flow diagram of an example process 200 for generating health event data for a temporal sequence. For convenience, the process 200 will be described as being performed by a system of one or more computers located at one or more locations. For example, a suitably programmed neural network training system, such as healthcare analysis system 100 of FIG. 1, may perform process 200.

The system receives an input temporal sequence (step 202). The temporal sequence includes data identifying each health event in each of the plurality of time steps. In some implementations, the temporal sequence is derived from an electronic medical record and includes data identifying each health event from the electronic medical record at each of the plurality of time steps. The health events in the sequence are ordered in chronological order, so that the most recently occurring health event becomes the health event at the last type step in the sequence.

The system processes the input temporal sequence using a recurrent neural network, eg, the recurrent neural network 110 of FIG. 1 to generate a neural network output for the input temporal sequence (step 204).

Depending on the implementation and the architecture of the recurrent neural network, the neural network output generated by the recurrent neural network by processing the input temporal sequence may include the following input scores, future condition scores, or network internal state. Can be.

The system generates health analysis data for the temporal sequence from the neural network output (step 206). As described above, the health analysis data depends on the type of neural network output generated by the recurrent neural network.

3 is a flow diagram of an example process 300 for generating health analysis data for a temporal sequence from subsequent input scores. For convenience, process 300 will be described as being performed by a system of one or more computers located at one or more locations. For example, a properly programmed neural network training system, such as healthcare analysis system 100 of FIG. 1, may perform process 300.

The system receives an input temporal sequence (step 302).

The system processes the input temporal sequence using the current neural network to generate the next input scores for the input temporal sequence (step 304). The recursive neural network includes one or more recurrent neural network layers and an output layer, where the output layer is generated by the recursive neural network layers for the time step, for each time step in the temporal sequence. Receive an internal state and generate a next set of input scores for the time step. The next set of input scores for the time step includes a respective score for each health event represented by the token in the vocabulary of tokens, where the next input score for a given health event is the health event in the temporal sequence. Denotes the probability of becoming the next health event, i.e., the possibility of becoming a health event at the next time step in the temporal sequence.

The next input scores for the input temporal sequence are the next input scores generated by the output layer for the last time step in the temporal sequence.

The system uses the following input scores to identify one or more highest-score health events (step 306). For example, the system may select a predetermined number of health events with the highest next input scores, or may select each health event with a next input score that is greater than a threshold value.

The system provides for presenting to the user data identifying the health events with the highest-scores, and optionally data characterizing the next input score for each highest-scores health event. (308)). Thus, the doctor or other user can view information about health events that are likely to be the next health events associated with the patient corresponding to the input temporal sequence.

4 is a flow diagram of an example process 400 for generating health event data for a temporal sequence from a network internal state. For convenience, process 400 will be described as being performed by a system of one or more computers located at one or more locations. For example, a properly programmed neural network training system, such as the neural network training system 100 of FIG. 1, may perform the process 400.

The system processes each of a set of temporal sequences using a recurrent neural network, eg, the recurrent neural network 110, to generate a network internal state for each time step of each temporal sequence (step (402)). Each temporal sequence in this set corresponds to a different patient, for example from a different electronic medical record. A recurrent neural network includes one or more recurrent neural network layers, and includes an output layer, or a set of logistic regression nodes, or both. In particular, the recurrent neural network may, for each time step in a given input temporal sequence, from the internal state generated by the neural network for the current time step, after future events, i.e., after the event at the current time step. It is trained to predict events that occur. For example, if the recurrent neural network includes an output layer, the recurrent neural network may be trained to predict the next event in the temporal sequence, that is, the event at the next time step after the current time step in the temporal sequence. . As another example, if the recursive neural network includes a set of logistic regression nodes, the recursive neural network may determine whether each of the set of events will occur within a specific period of time at the current time step in the temporal sequence. It may be trained to predict whether or not.

The system stores network internal states in an internal state repository, and associates each network internal state with data identifying the time step and temporal sequence that was targeted when the network internal state occurred (step 404). . For some implementations, for each temporal sequence, the system stores in the repository the network internal state generated by the system for each time step in the temporal sequence. For some other implementations, the system only stores a subset of network internal states in the repository, for example for those health events that are preceded by at least a threshold number of other health events in a temporal sequence. Store only network internal states.

The system receives an input temporal sequence of health events (step 406).

The system processes the input temporal sequence using the current neural network to determine the sequence internal state for the input temporal sequence (step 408). The sequence internal state for the input temporal sequence is the network internal state for the health event at the last time step in the sequence.

The system selects one or more network internal states from the internal state repository that are similar to sequence internal states (step 410). The system selects network internal states by calculating a similarity measure, for example, a cosine similarity measure, between the network internal states in the repository and the internal sequence state. For example, the system may select a predetermined number of network internal states with the greatest cosine similarity with the sequence internal state, or each network whose cosine similarity with the sequence internal state exceeds a threshold similarity. The internal state can be selected. In some implementations, the system uses other distance measures such as Euclidian distance, Hamming distance, etc. to determine the similarity between internal states. Similarly, the system can also normalize the internal states, and then calculate the distance between the normalized internal states.

The system provides for presenting the user with data identifying the temporal sequences of interest when similar network internal conditions occur (step 412). In particular, the system may, for any similar network internal state, occur within a temporal sequence of interest when a similar network internal state occurs, after a time step of that target when the network internal state occurs. Provide data identifying the events. Since a neural network that has generated both a sequence internal state and similar network internal states has been trained to predict future events from network internal states, and because similar network internal states are similar to sequence internal states, some network internal states When is generated, events occurring after the target time step are likely to predict future events in the input temporal sequence, that is, events occurring after the current last event in the input temporal sequence. In other words, from a targeted time step when a certain similar network internal condition occurs, the corresponding patient has a future that is similar to the future that the predictive neural network expects for the current patient corresponding to the input temporal sequence. It is expected by the Recurrent Neural Network to have. Thus, by viewing subsequent events from network internal conditions, the user, e.g., the physician, may experience events that may occur after the current last event in the input temporal sequence, ie the future that may occur for the current patient. Ideas of health events may be provided.

In some other implementations, the system also identifies data that identifies other health events in the temporal sequences to present to the user as part of the data identifying the temporal sequence of interest when a predetermined network internal condition occurs. to provide.

In some implementations, instead of providing data identifying temporal sequences for presentation to a user, the system calculates statistics from subsequent events in the temporal sequences and presents the calculated statistics to the user. To provide. For example, the system may include a portion of the temporal sequences that involved a particular health event, such as a heart attack or stroke, that occurs after a targeted time step when a similar internal network condition occurs. You can decide. In this case, the system may provide data identifying the proportion to present to the user, for example, "X% of patients will have futures similar to what the current patient has experienced in a particular health event. Expected to ".

In some implementations, instead of storing internal states in an internal state repository, the system re-calculates internal states for each other temporal sequence each time an input temporal sequence is received that is intended to be compared to other temporal sequences. (re-compute)

5 is a flowchart of an example process 500 for generating health event data for a temporal sequence from future condition scores. For convenience, process 500 will be described as being performed by a system of one or more computers located at one or more locations. For example, a suitably programmed neural network training system, such as the neural network training system 100 of FIG. 1, may perform the process 500.

The system receives an input temporal sequence (step 502).

The system processes the input temporal sequence using the recurrent neural network, eg, the recurrent neural network 110, to generate future condition scores for the input temporal sequence (step 504). Future condition scores include respective future condition scores for each of a predetermined set of conditions. The future condition score for a given condition indicates the likelihood that the condition is met within the specified time period of the event at the last time step of the input temporal sequence.

In some implementations, the recurrent neural network includes a set of one or more recurrent neural network layers and logistic regression nodes. Each logistic regression node generates, at each time step of the input temporal sequence, a future condition score for the corresponding condition from a predetermined set of conditions. A recursive neural network that includes logistic regression nodes that generate future condition scores is described in more detail below with reference to FIGS. In such implementations, the set for future condition scores generated by the current neural network for the last time step of the input temporal sequence is the set of future condition scores for the input temporal sequence.

In some other implementations, the recurrent neural network generates an next set of input scores for each time step of the input temporal sequence and includes an output layer that does not include logistic regression nodes. In these implementations, the system generates a number of possible temporal sequences, each temporal sequence comprising a specified number of additional time steps since the current last step in the temporal sequences and a possible health event in each of the additional time steps. . The system generates a number of possible temporal sequences by performing a beam search with a specific width for each of the additional time steps. The width of the beam search defines the number of highest scored events considered by the system at each of the future time steps. The system then determines, for each of the conditions met by the occurrence of one of the events for which future condition scores are to be generated, the ratio of possible temporal sequences comprising the event meeting the condition in one of the additional time steps of the sequence. Decide The system can then use the ratio as a future condition score for that condition. Optionally, the system can weight each occurrence of the event using the likelihood of occurrence of a possible temporal sequence in which the event occurred. The probability of occurrence of a possible temporal sequence may be, for example, the product of the next input scores for health events at each of the additional time steps in the sequence.

The system provides data identifying future condition scores for presentation to the user (step 506). For example, the system may provide data identifying each condition and a future condition score for each condition, or only provide data identifying one or more high score conditions for presentation to a user.

In some implementations, in addition to or in addition to providing data identifying future condition scores for presentation to a user, the system can determine the effect of processing on future condition scores and the system Data identifying the effect may be provided for presentation to the user.

6 is a flowchart of an example process 600 for determining the effect of adding an event to a temporal sequence for future condition scores. For convenience, process 600 will be described as being performed by a system of one or more computers located at one or more locations. For example, a properly programmed neural network training system, eg, neural network training system 100 of FIG. 1, may perform process 600.

The system receives an initial input temporal sequence (step 602).

The system determines future condition scores for the initial input temporal sequence (step 604). For example, the system may determine future condition scores for the initial input temporal sequence as described above with reference to FIG. 5.

The system receives data identifying additional health events from the user (step 606). For example, the additional health event may be a potential treatment that a doctor can prescribe to the patient.

The system generates a modified input temporal sequence by adding data identifying additional health events (eg, tokens representing health events) to the end of the initial input temporal sequence (step 608).

The system determines future condition scores for the modified input temporal sequence (step 610). For example, the system may determine future condition scores for the initial input temporal sequence as described above with reference to FIG. 5.

The system determines the change in future condition score caused by adding an additional health event to the input temporal sequence (step 612) and provides data identifying the change for presentation to the user (step 614). That is, the system calculates the difference between the future condition scores for the modified input temporal sequence and the corresponding future condition scores for the initial input temporal sequence and provides data identifying the differences for presentation to the user. . Thus, the physician can see the effects of potential treatments on the likelihood that certain conditions will be met in the future.

In some implementations, the system can automatically perform process 600 in response to a new event added to the temporal sequence. If a new event causes a condition's future condition score to increase or exceed the threshold, the system can automatically generate an alert to notify the user of the change. For example, a system administrator or other user may designate one or more specific conditions as undesirable. The system can then automatically perform the process 600 in response to the addition of a new event to the temporal sequence, and the future condition score for one of the undesirable conditions is above the threshold score or above the threshold increase. It will generate an alert to notify the user when increasing.

Further, in some implementations, the system can automatically generate multiple modified temporal sequences from the temporal sequence in response to receiving the temporal sequence, with each modified temporal sequence generating a different possible input health event in a temporal sequence. Add to Possible input health state events may be a subset of the health events represented by the token in the vocabulary, eg, some or all of the possible treatments represented by the tokens in the vocabulary. The system may then perform process 600 for each of the modified time sequences, and for any of the modified sequences, the future condition score for one or more undesirable conditions is reduced by a threshold. Determine whether to decrease more. In response to determining that for a given modified temporal sequence, the future condition score for undesirable conditions decreases above the threshold, the system identifies the health event that was added to the temporal sequence to generate the modified temporal sequence. Information to the user can be provided. Thus, the physician may have the opportunity to consider additional treatment methods that may reduce the likelihood that undesirable conditions will be met in the future.

7 illustrates an example recurrent neural network 700 configured to generate future condition scores. Recurrent neural network 700 is an example of a system implemented as computer programs on one or more computers at one or more locations, and the systems, components, and techniques described below may be implemented.

Recurrent neural network 700 receives input sequences comprising each input in each of a plurality of time steps, and for each of the time steps, obtains a respective future condition score for each condition in a predetermined set of events. Create The future condition score for a given condition at a given time step indicates the likelihood that the condition is met within a specified time period of input at the time step.

The recurrent neural network 700 includes one or more recurrent neural network layers 710, multiple logistic regression nodes 720A-720N, and optionally an output layer 740.

As shown in FIG. 1, for each time step, one or more concurrent neural network layers 710 receive input at the time step and aggregate the input to generate a network internal state for the time step. Processing

Each of the logistic regression nodes 720A- 720N corresponds to a respective condition from a set of predetermined conditions, and at each time step receives a network internal state for the time step and a future condition for the corresponding event. And process the network internal state according to current values of each set of parameters to generate a score. Thus, at each time step, each of the logistic regression nodes 720A- 720N generates a future condition score for each of the conditions within the predetermined set of conditions.

If the recurrent neural network 700 includes an output layer 740, the output layer 740 receives the network internal state during the time step and and each next input score for each possible input in the set of possible inputs. Configured to process the internal state to generate a. The next input score for a given possible input indicates the likelihood that the possible input is the next input of the input sequence, that is, the probability that it will appear immediately after the input at the current time step of the input sequence.

The inputs of the temporal sequence include inputs selected from tokens in a predetermined vocabulary representing a set of possible input events. Conditions in the set of predetermined conditions for which the recurrent neural network 700 generates future condition scores may include conditions that are satisfied by the occurrence of events that are not represented by tokens in the predetermined vocabulary. That is, the conditions within the predetermined set of conditions may be possible input events, events represented by tokens, or both, which may be included in the temporal sequences processed by the recurrent neural network 700. none. Thus, events in the set of events that satisfy any of the conditions in the set of predetermined conditions for which the recurrent neural network 700 generates future condition scores may overlap with the events represented by the tokens. The set of events may also include other events not in the set.

8 is a flow diagram of an example process 800 for generating future condition scores for a given time step. For convenience, the process 800 will be described as being performed by a system of one or more computers located at one or more locations. For example, a suitably programmed, redundant neural network, such as the redundant neural network of FIG. 7, may perform process 300.

The system receives an input for a time step, eg, a token representing a health event (step 802).

The system uses one or more of the concurrent neural network layers, for example, the selective neural network layers 710 of FIG. 7, to generate an in-network state for the concurrent neural network during the time step. Process (step 804). One or more neural network layers, for example, create an internal state of the network as described above with reference to FIG. 1.

The system processes the network internal state using each of the logistic regression nodes, eg, logistic regression nodes 720A-720N of FIG. 7, to generate a set of future condition scores (step 806). Each of the logistic regression nodes generates a future condition score for the corresponding condition by corresponding to each condition from a set of predetermined conditions and processing the internal state according to the current values of the set of parameters of the logistic regression node.

Optionally, the system also processes the network internal state using an output layer, eg, the output layer 740 of FIG. 7, to generate each next input score for each of the set of possible inputs. 808). The output layer generates each next input score by processing the network internal state according to the current values of the set of output layer parameters.

Process 800 may be performed on a neural network input for which the desired output (ie, the neural network output that should be generated by the system for input) is unknown. The system also sets a set of training sequences to train the system, that is, to determine the trained values for the parameters of the output layer, such as in recurrent neural network layers, logistic feedback nodes, and in some implementations. The process 800 can be performed on a set of inputs in which the inputs in E, the output that should be predicted by the system are known. In particular, process 800 may be performed repeatedly on inputs from a set of training sequences as part of a machine learning training technique for training a neural network, eg, back-propagation training technique over time. have. An exemplary training process is described in more detail below with reference to FIG. 9.

9 is a flow diagram of an example process 900 of training a recurrent neural network to generate future condition scores. For convenience, the process 900 will be described as being performed by a system of one or more computers located at one or more locations. For example, a properly programmed recurrent neural network, eg, the recurrent neural network 700 of FIG. 7, may perform the process 700.

The system obtains labeled training sequences (step 502). Each of the acquired training sequences is a sequence of inputs in each of a plurality of time steps. Each training sequence also includes, at each of the time steps, an indicator variable for each of the conditions in the set of predetermined conditions for which the current neural network generates future condition scores. The indicator variable for a given condition at a given time step indicates whether the condition has been met within a specified time period from the input of the time step. For example, the indicator variable may have a value of 1 if the condition is met and may have a value of 0 if the condition is not met. Thus, at each time step, the labeled training sequence includes an input for each condition and each indicator variable in the set of predetermined conditions.

In some implementations, the system receives training sequences already labeled with indicator variables. In some other implementations, the system generates labeled training sequences by calculating indicator variables for each of the conditions at each of the time steps. For example, the system may access, for a given input at a given time step of the training sequence, data that determines when the input occurred and identifies occurrences of events that satisfy conditions of a predetermined set of conditions. . The system then determines, for each condition, whether the condition has been met within the specified time at which the input occurred in the time step, and accordingly sets the value of the indicator variable for the event accordingly.

The system trains one or more literal neural network layers, logistic regression nodes, and optionally an output layer on labeled training sequences (step 504). In particular, the system determines the trained values of the parameters of the recurrent neural network layers, logistic regression nodes and output layers from the initial values of the parameters by performing multiple iterations of the machine learning training technique. As part of the training technique, the system minimizes or maximizes the objective function. If the system includes only logistic regression nodes and does not include an output layer, the objective function is for the given time step of the given training sequence, at the future condition score and time step generated by the logistic regression node for the time step. It depends on the error between the indicator variables for the corresponding conditions of. If the system also includes an output layer, the objective function also depends on the error between the next input scores generated by the output layer for the time step and the input at the next time step of the training sequence, for the time step. do.

As described above, the recurrent neural network 700 may process temporal sequences that include data identifying health events associated with the patient to generate future condition scores. However, the recurrent neural network 700 may include a sequence of temporal sequences that contain data identifying any type of temporal event, i.e., data that identifies aligned events as they occur over time. Can be trained to generate future condition scores for the temporal sequences of.

For example, the recurrent neural network 700 may identify transactions found in a user's financial statements, such as bank transactions that may appear on a bank statement, credit card transactions that may appear on a credit card statement, and the like. It can be trained to generate future condition scores for temporal sequences that include data. Future condition scores in this context are scores for conditions met by the various types of financial transactions being performed, scores for conditions met due to events that are not financial transactions of the kind indicated in the financial statements ( For example, tax audits) or both.

As another example, the recurrent neural network 700 may be trained to generate future condition scores for temporal sequences that include data identifying stock market transactions. In this context, temporal sequences may include buying and selling stocks of a single entity or all entities participating in the stock market.

As yet another example, the recurrent neural network 700 may include a data or a future sequence of temporal sequences that includes data identifying maintenance records for an electronic device, such as an aircraft, a vehicle, a data center component, or the like. It can be trained to generate a condition score. Future condition scores in this context are conditions that are satisfied with the occurrence of events that do not normally appear in maintenance records, as well as scores for conditions met by various types of maintenance related events (eg, Scores for errors in flight).

Embodiments of the subject matter and functional operations described herein include, but are not limited to, digital electronic circuitry, substantially implemented computer software or firmware, computer hardware, including the structures and structural equivalents disclosed herein, or combinations of one or more thereof. Can be implemented. Embodiments of the subject matter described in this specification are one or more computer programs, ie computer program instructions, executed by a data processing apparatus or encoded on an actual non-transitory program carrier for controlling the operation of the data processing apparatus. It can be implemented as one or more modules of the. Alternatively or additionally, the program instructions may be artificially generated and propagated to a signal propagated, for example, machine-generated electricity generated to encode information for transmission to an appropriate receiver device for execution by the data processing device, It can be encoded on an optical or electromagnetic signal. The computer storage medium may be a machine readable storage device, a machine readable storage substrate, a random or serial access memory device, or a combination of one or more thereof.

The term "data processing apparatus" includes all types of apparatus, devices and machines for data processing, including, for example, a programmable processor, a computer, or a plurality of processors or computers. The device may include special purpose logic circuits, eg, field programmable gate arrays (FPGAs) or application specific integrated circuits (ASICs). The apparatus may also, in addition to hardware, code that creates an execution environment for a computer program in question, such as processor firmware, protocol stacks, database management systems, operating systems, or combinations of one or more thereof. It may include.

A computer program (which may be referred to or described as a program, software, software application, module, software module, script, or code) may be any form of programming language, including compiled or interpreted or declarative or procedural languages. And may be distributed in any form including standalone programs, modules, components, subroutines, etc., or in other units suitable for use in a computing environment. A computer program may, but is not necessarily, correspond to a file in a file system. A program may be part of a file that holds one or more scripts stored in other programs or data, such as a markup language document, in a single file dedicated to that program, or part of one or more modules, subprograms, or code. Can be stored in a number of coordinated files, such as files that store them. The computer program may be distributed to run on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected to a communication network.

The processes and logic flows described herein may be performed by one or more programmable computers executing one or more computer programs to perform functions by manipulating input data and generating output. Processes and logic flows may also be performed by special purpose logic circuits such as field programmable gate arrays (FPGAs) or application specific integrated circuits (ASICs), and the apparatus may also be implemented in special purpose logic circuits.

A computer suitable for the execution of a computer program may be based, for example, on a general purpose or special purpose microprocessor or both, or any other kind of central processing unit. In general, the central processing unit will receive instructions and data from a read only memory or a random access memory or both. Essential elements of a computer are a central processing unit for performing or executing instructions and one or more memory devices for storing instructions and data. Generally, a computer is used to receive data from or transmit data to one or more mass storage devices for storing data (eg, magnetic, magneto-optical disks or optical disks). Or operatively combined for both receiving and transmitting. However, a computer may not need such devices. The computer may also include other devices such as mobile phones, personal digital assistants (PDAs), mobile audio or video players, game consoles, global positioning system (GPS) receivers or portable storage devices (e.g., USB ( universal serial bus).

Computer-readable media suitable for storing computer program instructions and data include, for example, semiconductor memory devices such as EPROM, EEPROM, and flash memory devices; Magnetic disks such as internal hard disks or removable disks; Magneto-optical disks; And all forms of non-volatile memory, media and memory devices, including CD ROM and DVD-ROM disks. The processor and memory may be supplemented by or included in special purpose logic circuits.

In order to provide for interaction with a user, embodiments of the subject matter described herein include a display device such as a cathode ray tube (CRT) or liquid crystal display (LCD) monitor for displaying information to a user, a keyboard, and It can be implemented on a computer with a pointing device, such as a mouse or trackball, that a user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user. For example, the feedback provided to the user can be any form of sensory feedback, such as visual feedback, auditory feedback, or tactile feedback; And input from the user may be received in any form, including acoustic, voice, or tactile input. The computer may also interact with the user by sending and receiving documents and devices used by the user. For example, the computer may interact with the user by sending a web page to a web browser on the user's client device in response to requests received from the web browser.

Embodiments of the subject matter described herein include, for example, a back end component as a data server, or include a middleware component such as an application server, or a client computer or user having a graphical user interface. To be implemented in a computing system that includes a front end component such as a web browser that can interact with the implementation of the described subject matter, or any combination of one or more such back end components, middleware components, or front end components. Can be. The components of the system may be interconnected by any form or medium of digital data communication, eg, a communication network. Examples of communication networks include local area networks (“LANs”) and wide area networks (“WANs”), such as the Internet.

The computing system can include clients and servers. Clients and servers are generally remote from each other and typically interact via a communication network. The relationship of client and server is caused by computer programs running on respective computers and having a client-server relationship with each other.

Although many specific implementations are included herein, these implementations are not to be construed as limitations on the scope of the invention or of what may be claimed, but rather are provided as a description of features that may be embodied in specific embodiments of specific inventions. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment can also be implemented individually or in any suitable subcombination in multiple embodiments. Moreover, although features may act as specific combinations and may be described as initially claimed as described above, one or more features from the claimed combination may be removed from the combination in some cases, and the claimed combination may be Subcombinations or variations of subcombinations.

Similarly, although the operations are shown in the drawings in a particular order, it is understood that in order to achieve desirable results, the operations are required to be performed in the specific order or sequential order shown or that all the operations shown are performed. Should not be. In certain situations, multitasking and parallel processing may be advantageous. In addition, the separation of the various system modules and components in the above embodiments should not be understood as requiring such separation in all embodiments, and the described program components and systems are generally integrated together in a single software product. It should be understood that it may be or may be packaged into multiple software products.

Specific embodiments of the invention have been described. Other embodiments are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. As one example, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In certain implementations, multitasking and parallel processing may be advantageous.

Claims (19)

A recurrent neural network system configured to process a temporal sequence comprising respective inputs in each of a plurality of time steps, wherein:
One or more concurrent neural network layers; And
One or more logistic regression nodes, the layers and nodes being trained on labeled training sequences,
The one or more concurrent neural network layers may be generated for each of the plurality of time steps:
Receive the input at the time step; And
Collectively process the input at the time step to create a network internal state for the time step;
Each of the logistic regression nodes corresponds to a respective condition of a set of predetermined conditions, each of the logistic regression nodes for each of the plurality of time steps:
Receive network internal state for the time step, and
Process a network internal state for the time step according to current values of the set of parameters of the logistic regression node to generate a future condition score for the corresponding condition for the time step,
And a future condition score for the corresponding condition indicates the likelihood that the corresponding condition is met within a specific time period of the input in the time step. The method of claim 1,
Further comprising an output layer, the output layer for each of the plurality of time steps:
Receive the network internal state for the time step;
Process the network internal state for the time step according to current values of the set of parameters of the output layer to generate each of a next input score for each of a plurality of possible next inputs,
Each of the next input scores for each of the plurality of possible next inputs indicates a likelihood that the possible next input will be an input at a next time step of the temporal sequence. The method according to claim 1 or 2,
Each of the one or more concurrent neural network layers may, for each of the plurality of time steps:
Receive a layer input for the time step; And
To generate a layer internal state for the recurrent neural network layer for the time step, according to the current values of the set of parameters for the recurrent neural network layer, for the recurrent neural network layer for a preceding time step. And a hierarchical internal state and hierarchical input for the time step. The method of claim 3,
Wherein said recurrent neural network system includes only one recurrent neural network layer, and said layer internal state for said recurrent neural network layer for said time step is said network internal state for said time step. Recurrent neural network system. The method of claim 3,
The recurrent neural network system includes a plurality of recurrent neural network layers arranged in a sequence, wherein the network internal state for the time step is a hierarchical internal state of the highest recurrent neural network layer of the sequence for the time step. Recurrent neural network system, characterized in that. The method of claim 3,
Wherein said recurrent neural network system comprises a plurality of recurrent neural network layers, wherein said network internal state for said time step is a combination of hierarchical internal states of said recurrent neural network layers for said time step. Recurrent Neural Network System. The method of claim 1,
Each input in the temporal sequence is selected from the set of possible inputs to the recurrent neural network system. The method of claim 7, wherein
The predetermined set of conditions includes conditions that are met if events that are not included in the set of possible inputs to the recurrent neural network system occur within a particular time period of the input in the time step. Recurrent neural network system. The method of claim 1,
Each of the inputs in the plurality of time steps is health-related data associated with a patient, and the conditions within the predetermined set of conditions occur when a health-related event occurs in association with the patient. Recurrent neural network system, characterized in that the conditions met. A method of processing a temporal sequence comprising respective inputs in each of a plurality of time steps using a recurrent neural network system comprising one or more logical neural network layers and one or more logistic regression nodes, the layers And nodes are trained on labeled training sequences, the method for each of the plurality of time steps:
Receiving the input at the time step;
Processing the input at the time step via one or more concurrent neural network layers to generate a network internal state for the time step; And
Processing the network internal state using each of the one or more logistic regression nodes,
Each of the logistic regression nodes corresponds to a respective condition of a set of predetermined conditions, and processing the network internal state using each of the one or more logistic regression nodes,
Processing the network internal state for the time step using the logistic regression node according to current values of the set of parameters of the logistic regression node to generate a future condition score for the corresponding condition for the time step. Including,
A future condition score for the corresponding condition indicates a likelihood that the corresponding condition is satisfied within a particular time period of the input in the time step. The method of claim 10,
The recurrent neural network system further comprises an output layer, the method further comprising: for each of the plurality of time steps:
Processing the network internal state for the time step using the output layer according to current values of the set of parameters of the output layer to generate each of a next input score for each of a plurality of possible next inputs. and,
Each next input score for each of the plurality of possible next inputs indicates the likelihood that the possible next input will be an input at a next time step of the temporal sequence. The method according to claim 10 or 11, wherein
Each of the one or more concurrent neural network layers may, for each of the plurality of time steps:
Receive a layer input for the time step; And
To generate a layer internal state for the recurrent neural network layer for the time step, according to the current values of the set of parameters for the recurrent neural network layer, for the recurrent neural network layer for a preceding time step. Processing a temporal sequence characterized by processing a layer internal state and a layer input for the time step. The method of claim 12,
Wherein said recurrent neural network system includes only one recurrent neural network layer, and said layer internal state for said recurrent neural network layer for said time step is said network internal state for said time step. A method of processing temporal sequences. The method of claim 12,
The recurrent neural network system includes a plurality of recurrent neural network layers arranged in a sequence, wherein the network internal state for the time step is a hierarchical internal state of the highest recurrent neural network layer of the sequence for the time step. And a temporal sequence characterized in that. The method of claim 12,
Wherein said recurrent neural network system comprises a plurality of recurrent neural network layers, wherein said network internal state for said time step is a combination of hierarchical internal states of said recurrent neural network layers for said time step. How to process a temporal sequence. The method of claim 10,
Each input in the temporal sequence is selected from a set of possible inputs to the recurrent neural network system. The method of claim 16,
The predetermined set of conditions includes conditions that are met if events that are not included in the set of possible inputs to the recurrent neural network system occur within a particular time period of the input in the time step. A method of processing temporal sequences. The method of claim 10,
Each of the inputs in the plurality of time steps is health-related data associated with a patient, and the conditions within the predetermined set of conditions occur when a health-related event occurs in association with the patient. A method for processing a temporal sequence characterized in that the conditions are met. A computer program stored on one or more non-transitory computer readable media, the computer program comprising instructions that when executed by one or more computers cause the one or more computers to perform: one or more concurrent neural network layers. 12. A method of processing a temporal sequence comprising a respective input at each of a plurality of time steps using a recurrent neural network system comprising a and a one or more logistic regression nodes, the operations according to claim 10 A computer program comprising a method.

Images